sdxl-unet-q8-gguf is an open-weight model from HYBRIDION, released under Apache License 2.0. Its published files total 10.1 GB.
SDXL UNet only, in four self-describing GGUF variants. Text encoders and VAE are not in these files — they come from any SDXL checkpoint. All four carry the same self-description KVs and load through the same patch/loader.
Model Card
By HYBRIDION, published under apache-2.0, revision a003b2d45296.
SDXL UNet only, in four self-describing GGUF variants. Text encoders and VAE are not in these files — they come from any SDXL checkpoint. All four carry the same self-description KVs and load through the same patch/loader. CPU and CPU-offload deployments should pick Q40 (fastest download, same runtime speed as the others once decoded). The file carries its own integration metadata as GGUF string KVs: Loaders read only the header to classify the file — no weight scan. Requires the SDXL-GGUF patch (config + loader + CUDA dequant dispatch): https://github.com/aeuphoraex/invokeai-sdxl-gguf The loader fetches text encoders + VAE from the embedded source-checkpoint path, falling back to any…
Read HYBRIDION's full model card
SDXL UNet GGUF (Q4_0 / Q8_0 / F16 / F32)
SDXL UNet only, in four self-describing GGUF variants. Text encoders and VAE are not in these files — they come from any SDXL checkpoint.
| file | quant | size | tensors | notes |
|---|---|---|---|---|
sdxl-unet-q4.gguf |
Q4_0 | 1.92 GB | 1628 Q4_0 + 52 F16 | smallest download; identical decoded compute as Q8 with eager-dequant loaders |
sdxl-unet-q8.gguf |
Q8_0 | 3.04 GB | 1628 Q8_0 + 52 F16 | recommended default; ~0.5% error, LoRA-safe |
sdxl-unet-f16.gguf |
F16 | 5.14 GB | 1680 F16 | bit-exact re-encode of the source UNet (still GGUF: lazy/streamed load) |
sdxl-unet-f32.gguf |
F32 | 10.27 GB | 1680 F32 | reference/analysis copy, no quant error at all |
All four carry the same self-description KVs and load through the same patch/loader. CPU and CPU-offload deployments should pick Q4_0 (fastest download, same runtime speed as the others once decoded).
Self-describing header
The file carries its own integration metadata as GGUF string KVs:
| KV | content |
|---|---|
sdxl.unet_config |
full UNet2DConditionModel architecture JSON |
sdxl.source_checkpoint |
path of the origin single-file checkpoint (provides TE + VAE) |
quantization.type |
Q8_0 |
Loaders read only the header to classify the file — no weight scan.
Use in InvokeAI (Model Manager → Install from URL)
Requires the SDXL-GGUF patch (config + loader + CUDA dequant dispatch): https://github.com/aeuphoraex/invokeai-sdxl-gguf
https://huggingface.co/AEUPH/sdxl-unet-q8-gguf/resolve/main/sdxl-unet-q8.gguf
# or: .../sdxl-unet-q4.gguf · .../sdxl-unet-f16.gguf · .../sdxl-unet-f32.gguf
The loader fetches text encoders + VAE from the embedded source-checkpoint path, falling back to any installed SDXL checkpoint if it doesn't exist.
Use in ComfyUI
With ComfyUI-GGUF:
UnetLoaderGGUF→ select any of the four.gguffilesDualCLIPLoader(clip-l + clip-g) +VAELoaderfrom any SDXL checkpoint- Standard KSampler / latents graph — see the
comfyui-workflow/folder in the companion repo.
Benchmarks (RTX 3090, 1024×1024, 30 steps, CFG 7)
| model | compute | VRAM after |
|---|---|---|
| SDXL checkpoint (fp8 storage) | 9.76 s | 15,412 MB |
| this file (Q8_0 GGUF) | 15.51 s | 5,172 MB |
74 % less VRAM; speed cost is per-op dequant (no fused kernel yet).
Conversion
Produced with sdxl_to_gguf.py from InvokeAI 6.14.1's bundled SDXL
checkpoint: every 2-D weight → Q8_0 via gguf.quants.quantize, written with
raw_dtype= so the GGUF writer accepts uint8 tensors; non-conforming conv
stays F16.
Integrity
ba6fca2f461df9d32ab24325cfb909b7 sdxl-unet-q8.gguf (md5, 3,040,509,728 B)
Identity and Version
- Repository
- AEUPH/sdxl-unet-q8-gguf
- Publisher
- HYBRIDION
- Task
- Not stated by the source
- Modality
- Other
- Library
- diffusers
- Parameters
- Not stated by the source
- Languages
- Not stated by the source
- Revision
- a003b2d45296e091d1023c4c89679eb27153c969
- First published
- 2026-09-30
- Last updated
- 2026-10-01
Files and Weights
5 files, 10.1 GB in total. The weights are 3 files totalling 10.1 GB in gguf.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| sdxl-unet-f16.gguf | Weights | 5.1 GB | d9f0d30fdec2 |
| sdxl-unet-q4.gguf | Weights | 1.9 GB | 5c43cf2fdbbd |
| sdxl-unet-q8.gguf | Weights | 3.0 GB | 3d67d31e0624 |
| README.md | Documentation | 3.0 KB | — |
| .gitattributes | Repository | 1.7 KB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 10.1 GB
Released by HYBRIDION through its official repository on Hugging Face. Read the license.
Built From
- Derived from stabilityai/stable-diffusion-xl-base-1.0
- Quantized from stabilityai/stable-diffusion-xl-base-1.0
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 10.1 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About sdxl-unet-q8-gguf
Can I use sdxl-unet-q8-gguf commercially?
Yes. sdxl-unet-q8-gguf is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.